US 7627801 B2 Abstract Methods and apparatus for encoding codewords which are particularly well suited for use with low density parity check (LDPC) codes and long codewords are described. The described methods allow encoding graph structures which are largely comprised of multiple identical copies of a much smaller graph. Copies of the smaller graph are subject to a controlled permutation operation to create the larger graph structure. The same controlled permutations are directly implemented to support bit passing between the replicated copies of the small graph. Bits corresponding to individual copies of the graph are stored in a memory and accessed in sets, one from each copy of the graph, using a SIMD read or write instruction. The graph permutation operation may be implemented by simply reordering bits, e.g., using a cyclic permutation operation, in each set of bits read out of a bit memory so that the bits are passed to processing circuits corresponding to different copies of the small graph.
Claims(20) 1. An apparatus for performing encoding operations, the apparatus comprising:
memory including a set of memory locations for storing L sets of Z-bit vectors, where Z is a positive integer greater than one and L is a positive integer;
a vector unit operation processor including an accumulator and output device for passing computed Z-bit vector to the said memory in response to operation instructions; and
a switching device coupled to the memory and to the vector unit operation processor, the switching device for passing a Z-bit vector between said memory and said vector unit operation processor in response to switch control information.
2. The apparatus of
3. The apparatus of
an ordering control module coupled to said memory for generating read and write indices; and
an operation control module coupled to said vector unit operation processor for generating unit operation instructions.
4. The apparatus of
5. The apparatus of
6. The apparatus of
7. The apparatus of
8. The apparatus of
9. The apparatus of
10. The apparatus of
11. The apparatus of
12. The apparatus of
13. A method of performing encoding operations, the method comprising:
storing L sets of Z-bit vectors in a memory device, where Z is a positive integer greater than one and L is a positive integer;
reading one of said sets of Z bit vectors from said stored L sets of Z bit vectors;
rotating the bits in said read one of said Z bit vectors; and
operating a vector unit processor to perform a plurality of combining operations to combine the bits of the rotated Z bit vector with a Z-bit vector stored in said vector unit processor to generate a new Z-bit vector.
14. The method of
storing said new Z bit vector in said memory device in the place of one of the stored L sets of Z bit vectors.
15. The method of
executing a set of stored machine executable instructions to control the rotation of the read Z bit vector.
16. The method of
using the executed set of stored machine executable instructions to determine which one of said sets of stored Z bit vectors is to be read from memory.
17. The method of
18. The method of
19. The method of
using the executed set of stored machine executable instructions to determine when one of said sets of stored Z bit vectors is to be read from memory.
20. The method of
using the executed set of stored machine executable instructions to determine which one of the stored L sets of Z bit vectors is to be replaced by storing the new Z bit vector in said memory device.
Description The present application is a continuation of U.S. patent application Ser. No. 10/618,325, filed on Jul. 11, 2003, now U.S Pat. No. 6,961,888, which claims the benefit of U.S. Provisional Patent Application Ser. No. 60/404,810 filed Aug. 20, 2002 titled “METHODS AND APPARATUS FOR ENCODING LDPC CODES” and U.S. Provisional Patent Application Ser. No. 60/450,245 filed Feb. 26, 2003 titled “PRODUCT LIFTINGS OF LOW-DENSITY PARITY-CHECK (LDPC) CODES” each of which is hereby expressly incorporated by reference. The present invention is directed to methods and apparatus for encoding data for the purpose of detecting and/or correcting errors in binary data, e.g., through the use of parity check codes such as low density parity check (LDPC) codes. Error correcting codes are ubiquitous in communications and data storage systems. Recently considerable interest has grown in a class of codes known as low-density parity-check (LDPC) codes. LDPC codes are often represented by bipartite graphs, called Tanner graphs, in which one set of nodes, the variable nodes, correspond to bits of the codeword and the other set of nodes, the constraint nodes, sometimes called check nodes, correspond to the set of parity-check constraints which define the code. Edges in the graph connect variable nodes to constraint nodes. A variable node and a constraint node are said to be neighbors if they are connected by an edge in the graph. For simplicity, we generally assume that a pair of nodes is connected by at most one edge. A bit sequence associated one-to-one with the variable nodes is a codeword of the code if and only if, for each constraint node, the bits neighboring the constraint (via their association with variable nodes) sum to zero modulo two, i.e., they comprise an even number of ones. In some cases a codeword may be punctured. This refers to the act of removing or puncturing certain bits from the codeword and not actually transmitting them. When encoding an LDPC code, however, bits which are to be punctured are still determined. Thus, puncturing has little or no impact on the encoding process. For this reason we will ignore the possibility of puncturing in the remainder of this application. The decoders and decoding algorithms used to decode LDPC codewords operate by exchanging messages within the graph along the edges and updating these messages by performing computations at the nodes based on the incoming messages. Such algorithms are generally referred to as message passing algorithms. Each variable node in the graph is initially provided with a soft bit, termed a received value, that indicates an estimate of the associated bit's value as determined by observations from, e.g., the communications channel. The encoding process, which is the focus of this application, also operates in part along the edges of the graph but the connection is less precise. The number of edges attached to a node, i.e., a variable node or constraint node, is referred to as the degree of the node. A regular graph or code is one for which all variable nodes have the same degree, j say, and all constraint nodes have the same degree, k say. In this case we say that the code is a (j,k) regular code. These codes were originally invented by Gallager (1961). In contrast to a “regular” code, an irregular code has constraint nodes and/or variable nodes of differing degrees. For example, some variable nodes may be of degree 4, others of degree 3 and still others of degree 2. While irregular codes can be more complicated to represent and/or implement, it has been shown that irregular LDPC codes can provide superior error correction/detection performance when compared to regular LDPC codes. While encoding efficiency and high data rates are important, for an encoding and/or decoding system to be practical for use in a wide range of devices, e.g., consumer devices, it is important that the encoders and/or decoders be capable of being implemented at reasonable cost. Accordingly, the ability to efficiently implement encoding/decoding schemes used for error correction and/or detection purposes, e.g., in terms of hardware costs, can be important. An exemplary bipartite graph While An alternative to the Tanner graph representation of LDPC codes is the parity check matrix representation such as that shown in In the case of a matrix representation, the codeword X which is to be transmitted can be represented as a vector The present invention is directed to methods and apparatus for performing encoding operations on binary data, e.g., multi-bit words. The methods and apparatus of the present invention allow for encoding of LDPC graphs that possess a certain hierarchical structure in which a full LDPC graph appears to be, in large part, made up of multiple copies, Z, e.g., of a Z times smaller graph. The Z graph copies may be identical. For purposes of explaining the invention, we will refer to the smaller graph as the projected graph. We refer to the Z parallel edges as vector edges, and Z parallel nodes as vector nodes. In U.S. patent application Ser. No. 09/975,331 titled “Methods and Apparatus for Performing LDPC Code Encoding and Decoding”, filed Oct. 10, 2001, which is hereby expressly incorporated by reference, we describe the benefits that such a structure lends to a decoder implementation. A key observation is that all operations may be done in parallel across all copies of the projected graph. The Z copies are not disjoint, however, they are combined to form one large graph, Z times larger than the projected graph. This is accomplished by interconnecting the Z copies of the projected graph in a controlled manner. Specifically, we allow the Z edges within a vector edge to undergo a permutation, or exchange, between copies of the projected graph as they go, e.g., from the variable node side to the constraint node side. In the vectorized message passing (decoding) process corresponding to the Z parallel projected graphs this exchange is implemented by permuting messages within a vector message as it is passed from one side of the vectorized graph to the other. The encoding process exploits the same idea, but the specification of the sequence of operations is somewhat different. In the encoding process all operations are performed on bit vectors rather than message vectors as in the decoding process. Consider indexing the projected LDPC graphs by 1, j, . . . , Z. In the strictly parallel graph variable nodes in graph j are connected only to constraint nodes in graph j. In accordance with the present invention, we take one vector edge, including one corresponding edge each from each graph copy, and allow a permutation within the Z edges, e.g., we permit the constraint nodes corresponding to the edges within the vector edge to be permuted, e.g., re-ordered. The re-ordering may be performed as rotations. For purposes of explaining the invention henceforth we will refer to the permutations, e.g., re-orderings, within the vector edges as rotations. A graph may be represented by storing information describing the projected graph and information describing the rotations. Alternatively, the description of the graph may be embodied as a circuit that implements a function describing the graph connectivity. Thus, in accordance with the present invention, a relatively large graph can be represented, e.g., described, using relatively little memory. Accordingly, the graph representation technique of the present invention facilitates parallel, e.g., vectorized, graph implementations. Furthermore, the graph representation techniques of the present invention can be used to support encoding of regular or irregular graphs, with or without state variables (punctured nodes). Note that normally all nodes belonging to a vector node will have the same degree, so degree information is required only for one projected graph. In various embodiments, the encoder is made programmable thereby allowing it to be programmed with multiple graph descriptions, e.g., as expressed in terms of a stored sequence of bit vector read/write and rotation information or in terms of an implemented function. Accordingly, the encoders of the present invention can be programmed to encode a large number of different codes, e.g., both regular and irregular. In some particular embodiments the encoder is used for a fixed graph or for fixed degrees. In such embodiments the graph description information may be preprogrammed or implicit. In such cases the encoder may be less flexible than the programmable embodiments but the resources required to support programmability are saved. Before presenting encoders for encoding large vectorized LDPC graphs, we will discuss general concepts and techniques relating to graph vectorization. The vectorization discussion will be followed by a presentation of exemplary vectorized LDPC encoders that embody the present invention. Vectorizing LDPC Graphs For purposes of gaining an understanding of vectorizing LDPC graphs consider a ‘small’ LDPC code with parity check matrix H. The small graph, in the context of a larger vectorized graph, will be referred to as the projected graph. Let Ψ denote a subset (usually a group) of Z×Z permutation matrices. We assume that the inverses of the permutations in Ψ are also in Ψ. Given the small, projected, graph we can form a Z-times larger LDPC graph by replacing each element of H with a Z×Z matrix. The 0 elements of H are replaced with the zero matrix, denoted 0. The 1 elements of H are each replaced with a matrix from Ψ. In this manner we ‘lift’ an LDPC graph to one Z times larger. The complexity of the representation comprises, roughly, the number of bits required to specify the permutation matrices, |E
Example: Lifting a small parity check matrix, the σ The subset Ψ can in general be chosen using various criteria. One of the main motivations for the above structure is to simplify hardware implementation of decoders and encoders. Therefore, it can be beneficial to restrict Ψ to permutations that can be efficiently implemented in hardware, e.g., in a switching network. Parallel switching network topologies is a well studied subject in connection with multiprocessor architectures and high speed communication switches. One practical example of a suitable architecture for the permutation subset Ψ is a class of multi-layer switching networks including, e.g., omega (perfect shuffle)/delta networks, log shifter networks, etc. These networks offer reasonable implementation complexity and sufficient richness for the subset Ψ. Additionally multi-layer switching networks scale well e.g., their complexity rises as N log N where N is the number of inputs to the network, which makes them especially suitable for massively parallel LDPC decoders. Alternatively, in decoders of the present invention with relatively low levels of parallelism and small Z the subset Ψ of permutations can be implemented in a single layer. An LDPC graph is said to have “multiple edges” if any pair of nodes is connected by more than one edge. A multiple edge is the set of edges connecting a pair of nodes that are connected by more than one edge. Although it is generally undesirable for an LDPC graph to have multiple edges, in many cases it may be necessary in the construction of vectorized graphs that the projected graph possesses multiple edges. One can extend the notion of a parity check matrix to allow the matrix entries to denote the number of edges connecting the associated pair of nodes. The codeword definition is still the same: the code is the set of 0,1 vectors x satisfying Hx=0 modulo 2. When vectorizing a projected graph with multiple edges, in accordance with the invention, each edge within the multiple edge is replaced with a permutation matrix from Ψ and these matrixes are added to yield the extended parity check matrix of the full code. Thus, a j>1 in the parity check matrix H of the projected graph will be ‘lifted’ to a sum σ The above described lifting appears to have one limitation. Under the above construction both the code length and the length of the encoded data unit must be multiples of Z. This apparent limitation is easily overcome, however. A description of the method used to overcome this limitation can be found in U.S. patent application Ser. No. 09/975,331 which is hereby expressly incorporated by reference and will not be repeated here. The invention lifts the encoding process analogously, replacing bit operations in the original algorithm to bit vector operations in the lifted algorithm. At one or more points in the encoding processing, after being read out of memory, the Z bit vectors are subject to a permutation operation, e.g., a re-ordering operation. The re-ordering operation may be a rotation operation, or rotation for short. These rotation operations generally correspond to the rotations associated to the vector edges which interconnect the Z copies of the projected graph to form the single large graph. In the case of encoding, however, some of the required rotations are apparent only after appropriate preprocessing of the LDPC representation. The rotation may be implemented using a simple switching device that connects, e.g., the bit memory to the bit vector processing unit and re-orders those bits as they pass from the memory to the bit vector processing unit. In such an exemplary embodiment, one of the bits in each bit vector read from memory is supplied to a corresponding one of the Z parallel processing units, within a bit vector processor, as determined by the rotation applied to the bit vector by the switching device. A rotation operation as implemented by the switching device may also or alternatively be applied to the bit vector prior to its being written into memory and after processing. The stored or computed description of the encoding process for the projected graph may include, e.g., information on the order in which bits in corresponding to a projected graph are to be read out of and/or written in to memory during encoding processing. The bits of the entire large graph are stored in multiple rows, each row corresponding to a different copy of the small graph, the rows being arranged to form columns of bits. Each column of bits represents a bit vector, which can be accessed as a single unit. The number of columns will typically be at least as large as the number of variable nodes in the projected graph, but often it will be larger, the additional columns being used for temporary storage in the encoding process. It is generally possible to decompose the encoding operation for lifted graphs into a sequence of elementary operations where each elementary operation consists of one of, e.g., reading a column of bits and rotating it, X-ORing that column bit-wise with some accumulated bit vector (possibly 0), and writing the result into some column in memory (usually additional rotation prior to writing is not required). As indicated above, to facilitate the encoding process it may be desirable or necessary to have more memory columns available then those required to store the codeword. In summary, the invention comprises the use of an encoding structure consisting of a switch to rotate bit vectors together with a bit-vector processor capable of performing the elementary operations described above and a control structure to control the sequence of operations performed, thereby specifying an encoding. Numerous additional advantages, features and aspects of the encoding techniques and encoders of the present invention will be apparent from the detailed description which follows. The encoding process for an LDPC code is a mapping from input information bits to an LDPC codeword. As discussed above, there are many possible forms this mapping can take. The present invention is directed towards a general purpose encoding device enabling fast parallel encoding of the class of LDPC codes supported by the decoder presented in application U.S. patent application Ser. No. 09/975,331. In that application, a certain structured class of LDPC codes was considered and a decoder architecture proposed for them. In this application certain features of the decoder architecture reappear as part of an encoder structure. For purposes of explaining the invention, we now describe a general purpose approach to encoding LDPC codes. The method is described in detail in a paper by Thomas J. Richardson and Ruediger L. Urbanke, titled “Efficient Encoding of Low Density Parity Check Codes” printed in the IEEE Trans. on Information Theory, pp. 638-656, Vol. 47, Number 2, February 2001. For purposes of discussion we assume that an m×n parity check matrix, has m<n and has rank m, that is, the rows are linearly independent. When this is not the case redundant rows can be removed without changing the code. We first describe certain operations which are part of the process of designing an encoder. It should be appreciated that this pre-processing computation is typically performed in software as part of code design and is not part of the actual implementation of the encoder. The first step in the design of an encoder according to our current method is to rearrange rows and columns to put the matrix H in approximate lower triangular form. Let x=(s,p An example is presented in The above description gives a method for encoding any LDPC code. It will be appreciated that many constructions of LDPC codes give rise to other natural encoding mechanisms, e.g. RA codes. The basic idea underlying our parallelized encoder is to take encoding methods for binary codes, such as described above, and “lift” them along with the parity check matrices into parallel an encoding engine for the “vectorized” LDPC codes. In a previously filed U.S. patent application Ser. No. 09/975,331 titled “Methods and Apparatus for Decoding LDPC Codes” which is hereby expressly incorporated by reference we described and motivated a structured “vectorized” class of LDPC graphs. The motivation there was to provide for a highly efficient decoder architecture. This application describes a corresponding architecture suitable for encoding the same class of codes. As in the decoder case, the advantages gained are that encoding operations may be performed efficiently and in parallel and the architecture allows the specification of the particular LDPC code to be programmable. We will now present a simple example of a small LDPC graph and its representation which will be used subsequently in explaining the invention. The discussion of the LDPC graph will be followed by a description of an LDPC encoder which can be used to encode the small graph. Matrix H Multiplication of a binary vector by a binary matrix can be decomposed into a sequence of simple operations. For example, consider multiplying a binary matrix U (m×n) with a binary vector v (n×1) in a hardware processor. We assume that, prior to multiplication, the vector v is available at some physical location, e.g. memory, starting at index s, and the result is to be stored at location starting at index t. Assume row i,i ∈[0,m−1] of matrix U has nonzero entries, i.e. 1's, at columns indexed as l With respect to the above allocation of memory The sequence instructions of We will now discuss in further detail the impact of vectorization on encoding techniques.-* Given a vectorized LDPC graph one can vectorize the encoding process as follows. The encoder operates as if it were encoding Z copies of the projected LDPC code synchronously and in parallel. Control of the encoding process corresponds to the projected LDPC graph and may be shared across the Z copies. Thus, we describe the encoder as operating on bit vectors, each vector having Z elements. One deviation from purely disjoint parallel encoding of the Z projected graphs is that bits are re-ordered within a bit vector during the encoding process. We refer to this re-ordering operation as a rotation. The rotation implements the permutation operations defined by Ψ. Because of the rotations, the processing paths of the Z copies of the projected graph mix, thereby linking them to form a single large graph. Control information which specifies the rotations is needed in addition to the control information required for the projected graph. Fortunately, the rotation control information can be specified using relatively little memory. While various permutations can be used for the rotations in accordance with the present invention, the use of cyclic permutations is particularly interesting because of the ease with which such permutations can be implemented. For simplicity we will now assume that Ψ comprises the group of cyclic permutations. In this case, our large LDPC graphs are constrained to have a quasi-cyclic structure. For purposes of this example, let N be the number of variable nodes in the graph and let M be the number of constraint nodes in the graph. First, we assume that both N and Mare multiples of Z, N=nZ and M=mZ where Z will denote the order of the cycle. Let us identify nodes through the use of a double index. Thus, variable node v The techniques of the present invention for representing a large graph using a much smaller graph representation and rotation information will now be explained further in reference to In accordance with the present invention, a larger graph can be generated by replicating, i.e., implementing multiple copies, of the small graph shown in Let us briefly discuss how to modify the Let us now consider the introduction of rotations into our example. This can be illustrated by replacing each of the 3×3 identity matrixes shown in Note that as a result of the vector edge permutation, operation, constraint node C We discussed above how to vectorize an encoder to encode Z parallel copies of the projected graph. By introducing switches into the message paths to perform rotations, we encode the LDPC code defined in The vector encoding process can be further appreciated by applying the general LDPC encoding procedure previously described in the present document. Instead of working on binary data, the encoder in accordance with the present invention works on a vector of Z bits, corresponding Z parallel copies of the bit in the projected graph. Parity check matrix H comprises entries of Z×Z all zero matrix or Z×Z cyclic permutation matrix represented by σ In constructing an encoder, preprocessing extracts and stores certain information. Matrix A Given the off-line pre-computed matrices, Similar to binary matrix multiplication decomposition described on page 21 of the present document and illustrated in Let us now consider how to decompose a multiplication of matrix U (m×n) comprising entries of Z×Z cyclic matrices or zero matrices with a vector v (n×1) of Z-bit data. Assume prior to multiplication, source data is held at locations s, s+1, . . . , s+n−1 in some memory of Z-bit data width; the result data is to be stored at locations t, . . . , t+m−1 in the same memory. Assume further that row i,i∈[0,m−1] of matrix U has nonzero entries, i.e. σ With respect to the above allocation of memory It will be apparent to those skilled in the field that the instructions listed in Table Some variations on the encoding methods and apparatus discussed above may result in reduced complexity in the case of some implementations. The following are some variations that may reduce the memory requirement for both the control memory -
- 1) Simplify the instruction representation:
- As described, an encoding instruction set is, in various embodiments, an ordered sequence of two basic instructions (0 a r b) and (1 a r b), which when executed produces the actual encoding. Such an instruction sequence may be generated by consecutively decomposing multiplications of some matrix with some vector into a sequence of basic instructions. Some exemplary decompositions include an overwhelming percentage of sub sequences of the following pattern: (0 a r
_{0 }b_{0}), (1 a r_{1 }b_{1}), . . . . (1 a r_{k }b_{k}). The repetition of a in this sub-sequence is redundant. This redundancy can be readily removed by modifying the basic instruction. Henceforth, we define two new instructions—(0 0 a) and (1 r a)—as follows: (1 r a) instructs the processor to read out the value at location a, left cyclic-shift it by r, and xor the value to the current value in an accumulator; (0 0 a) instructs the processor to write the current value in the accumulator to location a, and reset the value in the accumulator to zero. The transformation from the old instructions to the new instructions is clear: (0 a r b) is transformed to (1 r b), (0 0 a); and (1 a r b) is transformed to (1 0 a), (1 r b), (0 0 a). Following this rule, the exemplary sequence (0 a r_{0 }b_{0}), (1 a r_{1 }b_{1}), . . . , (1 a r_{k }b_{k}) is transformed to (1 r_{0 }b_{0}), (1 r_{1 }b_{1}), . . . , (1 r_{k }b_{k}), and (0 0 a), thus removing the redundancy. Transforming the instruction set in this manner can reduce the amount of memory required to implement control memory**1804**.
- As described, an encoding instruction set is, in various embodiments, an ordered sequence of two basic instructions (0 a r b) and (1 a r b), which when executed produces the actual encoding. Such an instruction sequence may be generated by consecutively decomposing multiplications of some matrix with some vector into a sequence of basic instructions. Some exemplary decompositions include an overwhelming percentage of sub sequences of the following pattern: (0 a r
- 2) Reduce the cardinality of the instruction set:
- When treating LDPC encoding as a sequence of matrices and vectors multiplications
**1600**, we can roughly divide the encoding process into three stages. In the first stage, we obtain T^{−1}As^{T }by first solving As^{T }then solving TZ=As^{T }in the second stage, we obtain p_{1}^{T}; and in the last stage given p_{1}^{T}, we obtain p_{2}^{T }by solving Tp_{2}^{T}=−As^{T}−Bp_{1}^{T}, which can be done efficiently using back-substitution. In the original form, matrices and vector multiplications in each stage are decomposed into an instruction subset. A sequential concatenation of those three subsets is the complete instruction set and the end of the instruction set implies the end of encoding process. However, sharing the instruction subset between the first stage and the last stage is possible and thus can reduce the cardinality of the instruction set. First, we note that T^{−1}As^{T }can be obtained by solving TP_{2}^{T}=−As^{T}−Bp_{1}^{T }if p_{1}^{T }is initialized to zero. Let us define the sequence of instructions to be the concatenation of the instruction subset for the last stage and for the second stage. So now encoding comprises 1) initialize p_{1}^{T }to be zero; 2) run the instruction subset for the last stage (obtain T^{−1}As^{T}) 3) run the instruction subset for the second stage (obtain p_{1}^{T}); 4) run the instruction subset for the last stage again (obtain p_{2}^{T}).
- When treating LDPC encoding as a sequence of matrices and vectors multiplications
- This instruction set sharing reduces the control memory
**1804**, and it will also reduce the encoding memory**1806**. It is because T^{−1}As^{T }is now saved at the location for p_{1}^{T }and there is no need in saving As^{T}.
- 1) Simplify the instruction representation:
Numerous additional variations on the encoding methods and apparatus of the present invention will be apparent to those skilled in the art in view of the above description of the invention. Such variations are to be considered within the scope of the invention. Patent Citations
Non-Patent Citations
Referenced by
Classifications
Legal Events
Rotate |